[PATCH V5 08/11] riscv: Support HAVE_IRQ_EXIT_ON_IRQ_STACK

Peter Zijlstra peterz at infradead.org
Tue Sep 20 00:34:19 PDT 2022


On Tue, Sep 20, 2022 at 09:27:51AM +0200, Peter Zijlstra wrote:
> On Tue, Sep 20, 2022 at 02:08:55PM +0800, Guo Ren wrote:
> > On Mon, Sep 19, 2022 at 9:45 PM Peter Zijlstra <peterz at infradead.org> wrote:
> > >
> > > On Sun, Sep 18, 2022 at 11:52:43AM -0400, guoren at kernel.org wrote:
> > >
> > > > +ENTRY(call_on_stack)
> > > > +     /* Create a frame record to save our ra and fp */
> > > > +     addi    sp, sp, -RISCV_SZPTR
> > > > +     REG_S   ra, (sp)
> > > > +     addi    sp, sp, -RISCV_SZPTR
> > > > +     REG_S   fp, (sp)
> > > > +
> > > > +     /* Save sp in fp */
> > > > +     move    fp, sp
> > > > +
> > > > +     /* Move to the new stack and call the function there */
> > > > +     li      a3, IRQ_STACK_SIZE
> > > > +     add     sp, a1, a3
> > > > +     jalr    a2
> > > > +
> > > > +     /*
> > > > +      * Restore sp from prev fp, and fp, ra from the frame
> > > > +      */
> > > > +     move    sp, fp
> > > > +     REG_L   fp, (sp)
> > > > +     addi    sp, sp, RISCV_SZPTR
> > > > +     REG_L   ra, (sp)
> > > > +     addi    sp, sp, RISCV_SZPTR
> > > > +     ret
> > > > +ENDPROC(call_on_stack)
> > >
> > > IIRC x86_64 moved away from a stack-switch function like this because it
> > > presents a convenient exploit gadget.
> > I found:
> > https://lore.kernel.org/all/20210204204903.350275743@linutronix.de/
> > 
> >   - The fact that the stack switching code ended up being an easy to find
> >     exploit gadget.
> > 
> > What's the exploit gadget? Do you have a ref link? Thx.
> 
> Sadly no, I do not. Kees might. But basically it boils down to this
> function taking both a stack pointer and a function pointer as
> arguments (@a1 and @a2 resp. if I'm not reading this wrong).
> 
> If an attacker can call this with arguments of its choice then it gains
> full control of subsequent execution.

If you inline it the hope is that the function pointers go away or at
least the encompassing function doesn't have quite such a 'convenient'
signature to hijack control flow.



More information about the linux-riscv mailing list